Insurance Companies Using AI in Customer Service: Examples

Insurance companies use AI in customer service in three main ways: an assistant that answers customers directly in chat or messaging apps, an assistant that helps service agents find answers faster, and tools that summarize cases and draft replies. Several large insurers have published case studies on these projects. This article lists the examples we could check against each company's own pages, what each one reported, and how far each number can be trusted.

Which insurance companies are using AI in customer service?

Eight insurers and insurance businesses have described their AI customer-service work in the public sources we read. They are spread across Germany, Taiwan, the United States, Switzerland, Asia, Spain, India and Croatia, and they build on different platforms: IBM, Google Cloud, Microsoft, AWS, Databricks, Infobip and the OpenAI API.

Every row comes from the company, or from a vendor describing that company's project. None of them is an independent audit. The dates also vary: Helvetia's release is from March 2023, SIGNAL IDUNA's post from March 2025 and AIA's story from January 2025, while several of the others carry no date. Check the date before you quote a figure, because these tools change fast.

What do insurers use AI for in customer service?

Most of these projects do one of three jobs. The job decides who the AI talks to, what it may say and which numbers matter, so it is the first thing to check when you read a case study.

AI that answers customers directly

Customer-facing assistants answer policy questions, handle simple changes and route harder cases to a person. Allianz Taiwan's Allie is the most detailed case in the sources. IBM's case study says Allie interprets complex insurance questions, responds to policy inquiries and makes policy changes, such as a policy loan or an address change, in one to two minutes, about the time it takes to dial into a call center. Allianz built her knowledge base from more than 400 real-life scripts drawn from a year of call-center requests.

LAQO, which Infobip describes as Croatia's first fully digital insurer, runs a WhatsApp assistant named Pavle. It was built on Azure OpenAI Service through Infobip's platform, and the build took two weeks. The team limited it to insurance claims and general information to avoid misleading answers, and complex queries go to a live agent.

John Hancock's contact-center bot, built with Microsoft's Azure Bot Service, answers routine questions on customer-facing websites and passes conversations to representatives in live chat. The project team described the aim as letting representatives focus on the most complex cases.

AI that helps agents answer faster

Knowledge assistants sit beside service staff and search the documents nobody can memorize. SIGNAL IDUNA's assistant is grounded in more than 2,000 internal documents covering more than 600 tariffs, so an agent can ask in plain language and get the relevant documents and an answer. Before the assistant, 27% of inquiries needed escalation to other departments or specialists, which the post names as a cause of delayed resolutions and higher costs.

Santalucía Seguros, a Spanish insurer, runs a retrieval-augmented assistant inside Microsoft Teams. It answers agents' questions about products, coverages and procedures, so that when a customer asks about coverage, the agent can answer in seconds.

InsuranceDekho is an Indian insurance marketplace that sells products from 49 insurers. Its assistant, built on Amazon Bedrock, answers questions about coverage, inclusions and exclusions from the policy documents, so agents and point-of-sale partners need not call subject-matter experts for each one.

AI that summarizes and drafts

The third job is the least visible to customers. AIA rolled out Microsoft's Copilot in Dynamics 365 Customer Service to its contact centers in Singapore and Thailand in 2024. Copilot summarizes case history, notes, emails and live chats, and it drafts email replies that representatives review before sending. The representative's work shifts from writing to checking.

What results have insurers published?

The numbers are real as published, but they measure different things by different methods. Put them side by side only with the method in view.

• SIGNAL IDUNA: the 30% figure is core processing time, meaning information search and response writing, in an experiment with 20 employees who handled requests with and without the assistant. The same post attributes a rise in case closure of about 24 percentage points, from 73% to almost 98%, to the assistant, and it does not describe how that rate was measured.

• Allianz Taiwan: the case study reports that nearly half of Allie's requests arrive outside call-center hours. Its 80% figure is the project's target, which the study reports was reached at launch: Allie covers 80% of the most frequent call-center requests. That is coverage of a request type, not an accuracy rate. The study also reports a customer rating of 4.5 out of 5 stars.

• AIA: AIA Thailand reports that Copilot conversation summaries cut wrap-up time, including review and tagging, by almost 50%.

• InsuranceDekho: an 80% decrease in the response time for queries about plan features, inclusions and exclusions.

• LAQO: 30% of customer queries are handled by the assistant, 90% are resolved within three to five messages, and agents resolve damage claims within 24 hours. These figures come from Infobip's customer story, which Infobip wrote about its own work with LAQO.

• Helvetia: no results. In March 2023 Helvetia said the Clara experiment would be evaluated with the Lucerne University of Applied Sciences and Arts before it decided on next steps. We found no published outcome in the sources we read, so check Helvetia's newsroom for later news.

Four questions make any of these numbers usable. What was compared: with and without the tool, or before and after. How many people or customers it covers: 20 employees is an experiment, while a whole contact center is a rollout. Who measured it: the company, a vendor or an independent body. And what the unit is: minutes, resolution, coverage or a rating. Ask those four questions of any AI customer-service claim, ours included.

Two gaps show up across these stories. None reports a measured error rate on real customer questions, and none reports what the project cost. A number without its denominator is a starting point for your own test, not a benchmark.

How does Tigerless use AI in customer service?

We build Lara, an AI platform for insurance, and we run it inside Tigerless Health, our insurance business. Lara is in production: it answers real customers at lara.tigerless.com ↗, where a customer can type a question or pick one of the sample questions. Lara is built against real plan documents, real deadlines and real customers, not as a demo.

Lara handles coverage questions, plan comparisons, enrollment, billing and claims, inside a licensed agency. It answers from the actual plan document rather than a generic definition. For a claim, it lists what is needed, in what order and by when. Read more on Lara's product page ↗.

Lara learns in production, from the questions real customers ask rather than from a synthetic evaluation set. The product page invites visitors to test it the hard way: "The fastest way to evaluate Lara is to ask her something hard." Try a question with a real deadline or a real plan in it, then check the answer against the plan document.

Lara does not sell a policy, approve a claim or give medical advice. It does the reading, the comparing and the explaining, and when a question needs a licensed agent, it brings one in. The product page says so as well: Lara is AI and can make mistakes, so double-check responses.

Two facts keep this comparison honest. First, Tigerless AI is not the insurer. The licensed agency is Tigerless Insurance Agency LLC, which sells plans at tigerless.com ↗, and plan facts such as carriers, prices, states and eligibility come only from that site. Our health business is described on Tigerless Health ↗. Second, we publish no price for Lara and no accuracy or benchmark score for it, so there is no Lara figure to set beside the examples above.

When is an AI assistant the wrong tool for an insurance customer?

An AI assistant is the wrong tool when the answer binds the insurer, when it depends on medical judgment, or when a deadline cannot slip. Every example above keeps a person in the loop somewhere. LAQO hands complex queries to live agents, John Hancock's bot passes conversations to representatives, and Allianz's address change ends with a one-time-password confirmation. Helvetia put it plainly in 2023: its chatbot "is in no way a substitute for an individual advisory discussion with a specialist."

Go to the insurer's own channel or a licensed agent when:

• you are about to buy a policy, change coverage or accept a settlement, because those are binding decisions;

• the question is medical, which an assistant is not built to answer;

• you plan to appeal a denied claim and the decision needs a licensed person to review it;

• the plan document does not answer your question.

Comparing plans is where an assistant helps most, but the plan you sign is a binding decision. Put the comparison in front of a licensed agent before you enroll.

Data is the other question. Before you share health details with any AI assistant, ask where the chat is stored and who can read it. Allianz's case study says it kept customer data on its own premises to meet regulatory requirements. Engineering teams that route health data through prompts, logs or memory can use phi-boundary-gate ↗, our open-source project in Tigerless Labs ↗, to report where candidate identifiers move. It treats every hit as a candidate for human review, and it makes no HIPAA-compliance claim.

We apply the same rule to our own AI-written blog drafts: a person reads each one before it goes live, as our review process ↗ describes.

Conclusion

Insurers use AI in customer service to answer customers, to help agents and to summarize and draft. The published results show real gains in time and resolution, but they come from insurers, vendors and cloud providers describing their own projects, not from independent audits. The best AI for an insurance customer answers from the actual plan document, says what it cannot do and hands any binding decision to a licensed person.

Use this rule: if a question changes what you pay, what is covered or what is paid out, a licensed person should confirm it. If the question is what your policy says, an assistant can read it with you. Your next step is small: open lara.tigerless.com ↗, ask about a plan you already know well, and check the answer against the plan document before you rely on it. Then put the insurer's published numbers beside yours and ask how each one was measured.